Implicit Neural Representation (INR) has emerged as a powerful approach for learning Signed Distance Function (SDF) in unsupervised 3D reconstruction from point clouds. However, current methods frequently suffer from over-permissiveness toward global viscosity solutions, resulting in spurious local minima that degrade both SDF optimization and final reconstruction quality. To overcome these limitations, we propose a novel loss function named Geometric-Differential Regularization (GDR) with two complementary components: Gradient-Projection regularization to suppress undesirable extrema formation, and Hessian-Degeneration regularization to enhance surface-proximal learning through higher-order geometric constraints. Furthermore, we develop a sampling strategy that dynamically adjusts constraints based on local geometric features of surface points. We validate our method through both 2D SDF learning visualizations and 3D reconstruction experiments using noisy point clouds, demonstrating significant improvements over existing approaches.